wiki-authoring

name	wiki-authoring
description	Use when creating or updating Hermes memory wiki content from verified context while preserving vault structure and validating the resulting pages.
version	0.1.0
author	Nous Research
license	MIT
metadata	{"tools":["wiki_get","wiki_apply","wiki_ingest","wiki_compile","wiki_lint"]}

Overview

Author OpenClaw-compatible memory wiki content by searching/reading existing context, ingesting source material when needed, applying the smallest useful typed mutation, compiling derived outputs, linting the vault, and refreshing search indexes when a tool workflow allows it. Prefer precise source-backed updates over broad rewrites. The tools are deterministic: there are no hidden tool-layer LLM calls, so the agent/LLM chooses content, claims, and evidence references explicitly before submitting tool operations.

When to Use

• A user asks to add a new entity, concept, claim, source, report note, or synthesis page.
• Existing wiki content needs a correction, clarification, or additional source-backed note.
• A draft update needs validation against the OpenClaw-compatible schema.
• A recent edit should be compiled and checked for wiki consistency.

Schema Essentials

Directory taxonomy defines broad page kind:

• entities/ pages use pageType: entity.
• concepts/ pages use pageType: concept.
• syntheses/ pages use pageType: synthesis.
• sources/ pages use pageType: source.
• reports/ pages use pageType: report.

Use entityType, not pageType, for entity subtypes. A person page is an entity page with entityType: person; organizations, projects, systems, and products are also subtypes under pageType: entity.

Required discipline:

• Keep id, title, and pageType in frontmatter.
• Put sourceIds on non-source/non-report pages so claims can be traced to source pages.
• Prefer structured claims over unsupported prose-only assertions.
• Give each claim a stable id, clear text, status, confidence, and updatedAt when known.
• Add evidence entries for source-backed claims.

Evidence format example:

claims:
  - id: claim.ada.role
    text: Ada is the project lead for the memory wiki work.
    status: active
    confidence: 0.9
    evidence:
      - kind: source
        sourceId: source.project-notes
        path: sources/project-notes.md
        lines: "12-18"
        weight: 1
        note: Project notes identify the role.
        confidence: 0.9
        privacyTier: standard
        updatedAt: "2026-05-28T00:00:00Z"
    updatedAt: "2026-05-28T00:00:00Z"

Authoring Workflow

1. Use wiki_search and wiki_get before editing to read the target page, nearby pages, matching aliases, source pages, or nearest related context.
2. Check for an existing entity/concept/source to avoid duplicate pages and preserve human notes.
3. Use wiki_ingest for source pages from local-file, conversation-summary, or text input before citing them. Conversation summaries are source pages; cite the returned source ID in sourceIds.
4. Prepare a minimal content change that matches the directory taxonomy and schema. The LLM/agent supplies the claim text and evidence; tools only validate and write deterministic files.
5. Use wiki_apply upsert_entity for typed entity pages and wiki_apply upsert_concept for typed concept pages. Include source-backed sourceIds, and include structured claims/evidence when making assertions.
6. Use wiki_apply create_synthesis for cross-source synthesis and wiki_apply update_metadata for targeted metadata changes.
7. Run wiki_compile so generated indexes, reports, and cache files reflect the change.
8. Run wiki_lint to catch malformed links, metadata issues, provenance gaps, or structural problems.
9. If search freshness matters after the edit, use the maintenance workflow to run wiki_reindex before relying on search.
10. If lint reports errors, inspect the affected page with wiki_get, then apply a focused fix with wiki_apply and repeat the compile/lint/reindex verification loop.

Pitfalls

• Do not invent details that are not supported by the conversation, source pages, or retrieved wiki context.
• Do not overwrite large sections when a targeted wiki_apply update is enough.
• Do not create arbitrary freeform Markdown pages; use wiki_ingest for sources and typed wiki_apply operations for authored pages.
• Do not assume the tool generated claims for you. The LLM/agent is responsible for the content and claims it submits.
• Do not treat a person as its own broad page kind; use pageType: entity plus entityType: person.
• Do not omit sourceIds from entity, concept, or synthesis pages unless the page is intentionally unsourced and the limitation is explicit.
• Do not claim the edit is ready until wiki_compile and wiki_lint have run after the change.

Verification Checklist

• Target content was inspected with wiki_search and wiki_get before editing.
• New source material was captured with wiki_ingest when applicable (local-file, conversation-summary, or text), and the update was applied through wiki_apply upsert_entity, wiki_apply upsert_concept, wiki_apply create_synthesis, or wiki_apply update_metadata.
• Frontmatter uses the correct broad pageType and any entity subtype uses entityType.
• Claims have source-backed evidence when possible and non-source/non-report pages list sourceIds.
• wiki_compile completed after the authoring change, wiki_lint was reviewed, and wiki_reindex was run when search freshness matters.
• wiki_lint reports the vault is valid or only has clearly explained non-blocking warnings.